* Your assessment is very important for improving the workof artificial intelligence, which forms the content of this project
Download Artificial Neural Networks Introduction to connectionism
Donald O. Hebb wikipedia , lookup
Neurotransmitter wikipedia , lookup
State-dependent memory wikipedia , lookup
Activity-dependent plasticity wikipedia , lookup
Nonsynaptic plasticity wikipedia , lookup
Neurophilosophy wikipedia , lookup
Stimulus (physiology) wikipedia , lookup
Cognitive neuroscience wikipedia , lookup
Neuroeconomics wikipedia , lookup
Feature detection (nervous system) wikipedia , lookup
Neural oscillation wikipedia , lookup
Single-unit recording wikipedia , lookup
Catastrophic interference wikipedia , lookup
Neural coding wikipedia , lookup
Artificial intelligence wikipedia , lookup
Optogenetics wikipedia , lookup
Cognitive science wikipedia , lookup
Embodied cognitive science wikipedia , lookup
Synaptogenesis wikipedia , lookup
Artificial general intelligence wikipedia , lookup
Central pattern generator wikipedia , lookup
Biological neuron model wikipedia , lookup
Neural modeling fields wikipedia , lookup
Convolutional neural network wikipedia , lookup
Neuroanatomy wikipedia , lookup
History of artificial intelligence wikipedia , lookup
Channelrhodopsin wikipedia , lookup
Holonomic brain theory wikipedia , lookup
Synaptic gating wikipedia , lookup
Neuropsychopharmacology wikipedia , lookup
Neural engineering wikipedia , lookup
Chemical synapse wikipedia , lookup
Artificial neural network wikipedia , lookup
Neural binding wikipedia , lookup
Metastability in the brain wikipedia , lookup
Development of the nervous system wikipedia , lookup
Recurrent neural network wikipedia , lookup
Introduction to connectionism
Artificial Neural Networks
Connectionism – theory of information processing, inspired by
biology (brains). It's based on Artificial Neural Networks (ANNs).
Introduction
It has two goals:
theoretical foundations of cognitive science (modeling mental
processes)
- contrasting with symbolic approaches
- features: parallelism, robustness, learning from experience,...
Igor Farkaš
Department of Applied Informatics
Comenius University in Bratislava
© Igor Farkas
ANN: Introduction
- tasks: pattern recognition, classification, associative memory, time series
prediction, dimensionality reduction, data visualization, ...
1
© Igor Farkas
Brain = highly complex, non-linear and parallel
information processing system (“computer”)
11
composed of ~ 10 neurons (basic information-processing
elements), connected via ~ 1014 synapses
on certain tasks it is much faster than supercomputers of
today, even though neurons are very slow (~ ms)
mostly prewired at birth, but very plastic throughout life
importance of learning: involves 3 mechanisms
© Igor Farkas
modification of existing synapses, generation of new synapses and/
or new neural cells
ANN: Introduction
ANN: Introduction
2
Structural organization of levels in the brain
Human brain
applications in practical problems
3
Central nervous system
Interregional circuits (columns, topographic maps)
Local circuits
(~1 mm)
Neurons
(~100 m)
Dendritic trees
Neural microcircuits (~ m)
Synapses
Molecules
© Igor Farkas
ANN: Introduction
4
1 mm3
Synapse
105 neurons
3 km axons
Synapse maintains the interaction between neurons.
Presynaptic process releases a neurotransmitter, which diffuses across
the synaptic junction between neurons and then acts on a postsynaptic
process.
Synapse mediates electrical-chemical-electrical signal conversion.
Effect on a postsynaptic neuron can be either excitatory or inhibitory.
n
From S.R. y Cajal: Texture of the Nervous System of
Man and the Vertebrates (illustrates the diversity of
neuronal morphologies in the auditory cortex).
© Igor Farkas
ANN: Introduction
5
© Igor Farkas
6
Typical artificial neuron model
Action potential
If a neuron is made to “fire”,
generated action potential (AP)
traverses along the axon,
uninhibited.
1. receives signals from other neurons (or sensors)
2. processes (integrates) incoming signals
3. sends the processed signal to other neurons (or muscles)
Generation of AP (neuron “firing”)
requires that membrane potential
exceed the excitation threshold.
Deterministic model
y = f (i wixi - )
This usually requires one of the
two conditions:
1. temporal summation - arrival of
multiple impulses in time at the
same synapse, or
2. spatial summation - impulses
arrive simultaneously at sufficient
number of different synapses.
© Igor Farkas
ANN: Introduction
ANN: Introduction
Stochastic model P(s=+1) = 1/(1+exp(-i wixi + ))
7
© Igor Farkas
ANN: Introduction
8
Useful features of artificial neural networks
Nonlinearity (of processing units)
Input-output mapping (nonparametric statistical inference)
Adaptivity (parameter tuning)
Evidential response (degree of 'confidence')
Contextual information ( thank to connectivity)
Fault tolerance (graceful degradation)
VLSI implementability
Neurobiological analogy
Importance of environment
Neural network architectures
Single-layer feedforward NNs
Multi-layer feedforward NNs
Have hidden layers
Partial connectivity
=
Recurrent networks
Associative memories
Input-output mapping networks
Layer of nodes
(Čerňanský, 2006)
© Igor Farkas
ANN: Introduction
9
© Igor Farkas
Aristoteles (400 BC) – introduced concepts of memory, and of
connectionism
Spencer (1855) – separated psychology from philosophy, postulated that
“neural states affect psychological states”, knowledge is in connections.
James (1890) – model of associative memory, “law of neural habit”
Thorndike (1932) – distinguishes sub-symbolic view on neural associations,
formulated two laws of adaptation: “the law of effect” and “the law of
exercise”.
McCullogh & Pitts (1943) – neural networks with threshold units
classical connectionism (until 1940s)
within philosophy, psychology
old connectionism (1950s-1970s) – birth of computer era
beginning of theory of artificial neural networks
linked to cognitive science revolution
new connectionism (from 1986)
parallel distributed processing subsymbolic processing
multi-layer NN models (incl. recurrent)
1. Logical neuron can simulate any linearly separable Boolean function.
n
2. Any Boolean function {0,1} {0,1} can be simulated by a 2-layer NN
with logical units.
even newer connectionism (late 1990s)
© Igor Farkas
Minsky (1957) reformulated their results to comprehensible form.
multilayer generative models (probabilistic approach)
ANN: Introduction
10
Classical connectionism
Brief history of connectionism
ANN: Introduction
11
© Igor Farkas
ANN: Introduction
12
Knowledge representation in AI
Knowledge refers to stored information or models used by a
person or machine to interpret, predict and appropriately
respond to the outside world. (Fischler & Firschein, 1987)
Primary characteristics of KR:
What information is made explicit
How the information is physically encoded for use
KR is goal oriented: In “intelligent” machines, a good solution
depends on a good KR.
ANNs are a special class of intelligent machines.
Knowledge in ANN is distributed in free parameters.
© Igor Farkas
ANN: Introduction
Rules for knowledge representation in ANN
1. Similar inputs from similar classes should usually produce
similar representations inside the NN, and should hence be
classified as belonging to the same category (measures of
similarity: e.g. Euclidean dist., Mahalanobis dist).
2. Items to be categorized as separate classes should be given
widely different representations in the NN.
3. If a particular feature is important, then there should be a
large number of neurons involved in the representation of that
item in the NN.
4. Prior information and invariances should be built into the
design of a NN.
13
© Igor Farkas
14
How about ANN as cognitive models?
Artificial Intelligence and ANNs
AI system must be able to:
1.
2.
3.
ANN: Introduction
Level of explanation:
store knowledge,
apply the knowledge
acquire new knowledge through experience.
Processing style:
Key components of AI system:
parallel, distributed
Representational structure:
- representation – symbolic (declarative, procedural)
subsymbolic (less transparent than symbolic)
representing structure is a challenge for NN
in NN: mostly subsymbolic
- reasoning – ability to solve problems (searching)
in NN: straightforward (propagation of activation)
- learning – inductive, deductive
in NN usually: training – inductive (trajectory in weight space),
testing - deductive
© Igor Farkas
ANN: Introduction
15
© Igor Farkas
ANN: Introduction
16
Learning in NN
Learning paradigms:
Pattern association
Pattern recognition / classification
Function approximation
Learning rule types:
Feature extraction
Error-correction
Control
Memory-based (e.g. k-NN)
Filtering
Hebbian
Competitive
Auto-associative memory
Boltzmann (stochastic)
Prediction
supervised (with teacher)
reinforcement learning (partial feedback from environment)
Learning tasks
© Igor Farkas
unsupervised (self-organized)
ANN: Introduction
17
© Igor Farkas
ANN: Introduction
18